Image Steganography Based on Three-Channel Deep Fusion Technology

Literature Overview

This paper by Liu Lianshan and Huang Yu from Shandong University of Science and Technology, published in Information Security Research (Vol. 11, No. 3, 2025, pp. 257–264), presents an image steganography method based on three-channel deep fusion technology. The work was supported by the Shandong Provincial Natural Science Foundation and the Shandong Provincial Key R&D Program (Soft Science). The method employs a U-Net-based main channel for cover image feature extraction with residual blocks, and lower and middle channels for secret image feature extraction, with cross-fusion between channels to improve steganographic image quality and embedding capacity.

Core Technical Points

The proposed method achieves a peak signal-to-noise ratio (PSNR) of 41.15 dB at an embedding capacity of 24 bits per pixel (bpp), demonstrating high invisibility and large embedding capacity. The three-channel architecture consists of:

• Main channel: U-Net structure with residual blocks for extracting cover image features.
• Lower channel: Extracts features from the secret image at a coarse level.
• Middle channel: Extracts features from the secret image at an intermediate level.

The cross-fusion strategy involves fusing features from layers 1 and 3 of the lower channel into the corresponding layers of the middle channel, and then fusing features from layers 2 and 4 of the middle channel into the corresponding layers of the main channel. This hierarchical fusion progressively integrates secret image information into the cover image representation.

Performance Comparison

Architecture and Fusion Analysis

The U-Net architecture provides a strong foundation for image-to-image translation tasks, with its encoder-decoder structure and skip connections preserving spatial information. The addition of residual blocks enhances the network's ability to learn complex feature mappings while mitigating the vanishing gradient problem in deep networks. The skip connections in U-Net are particularly valuable for steganography, as they help preserve the structural details of the cover image while allowing the network to learn how to embed secret information.

The three-channel fusion strategy is a key innovation. By extracting secret image features at multiple scales (lower and middle channels) and progressively fusing them into the main channel, the method captures both coarse and fine-grained secret information. This multi-scale approach is superior to single-channel methods that may miss either high-frequency details or low-frequency structural information in the secret image.

Integration with Engineering Practice

While this paper belongs to the domain of information security, the engineering methodology has parallels in steel pipe and fitting quality control. The concept of multi-scale feature extraction and fusion is analogous to the multi-scale inspection approach used in non-destructive testing (NDT) of welded pipes. Just as the steganography method extracts features at different scales to improve detection capability, NDT techniques such as phased array ultrasonic testing (PAUT) and time-of-flight diffraction (TOFD) examine welds at different spatial resolutions to detect defects of varying sizes.

The emphasis on invisibility in steganography—ensuring that the embedded information does not degrade the cover image quality—is conceptually similar to the principle of minimal intervention in NDT. The goal in both cases is to extract or embed information without significantly altering the original. In NDT, this means using non-destructive methods that do not damage the test article; in steganography, it means embedding data without degrading image quality.

Key Questions and Reflections

The 24 bpp embedding capacity at 41.15 dB PSNR represents a strong performance, but questions remain about robustness against various attacks. The paper does not address the steganographic strength against steganalysis attacks, compression, or noise addition. In practical applications, the steganographic image must survive transmission through communication channels that may introduce compression artifacts or noise, which could degrade the recoverability of the embedded secret information.

Another consideration is the computational complexity of the three-channel architecture. The method requires processing three separate channels with cross-fusion, which increases computational cost compared to single-channel approaches. For real-time applications or resource-constrained environments, this computational overhead may be a limiting factor.

Study Insights and Implications

The three-channel deep fusion approach demonstrates that hierarchical feature extraction and progressive fusion can significantly improve steganographic performance. The combination of U-Net architecture with residual blocks and multi-channel fusion provides a powerful framework for balancing embedding capacity and invisibility. For engineers in adjacent fields, the key lesson is that multi-scale feature processing and progressive integration of information can improve performance in complex signal processing tasks, whether in information security, non-destructive testing, or other engineering applications. The method's ability to achieve high embedding capacity while maintaining good image quality represents a significant advance in the field of digital steganography.